基于深度学习的COVID-19智能诊断系统

Intelligent Diagnosis System for COVID-19 Based on Deep Learning

摘要:

本研究旨在探索运用深度学习的方法辅助医生利用胸部X光片进行COVID-19诊断的可行性和准确性。首先利用公开的COVID-QU-Ex Dataset训练集训练一个UNet分割模型，实现肺部ROI区域的自动分割。其次完成对该公共数据集肺部区域的自动提取预处理。再次利用预处理后的三分类影像数据（新冠肺炎、其它肺炎、正常）采用迁移学习的方式训练了一个分类模型MBCA-COVIDNET，该模型以MobileNetV2作为骨干网络，并在其中加入坐标注意力机制（CA）。最后利用训练好的模型和Hugging Face开源软件搭建了一套方便医生使用的COVID-19智能辅助诊断系统。该模型在COVID-QU-Ex Dataset测试集上取得了高达97.98%的准确率，而该模型的参数量和MACs仅有2.23M和0.33G，易于在硬件设备上进行部署。该智能诊断系统能够很好的辅助医生进行基于胸片的COVID-19诊断，提升诊断的准确率以及诊断效率。

Abstract:

This study aims to explore the feasibility and accuracy of using deep learning to assist doctors in the diagnosis of COVID-19 by chest X-ray. Firstly, a UNet segmentation model was trained using the open COVID-QU-Ex Dataset training set to realize the automatic segmentation of lung ROI region. Secondly, the automatic extraction preprocessing of the lung region of the public dataset is completed. Thirdly, a classification model MBCA-COVIDNET is trained by using the preprocessed three classification image data (COVID-19, other pneumonia and normal) in the way of transfer learning. This model takes MobileNetV2 as the backbone network and adds the coordinate attention mechanism (CA) to it. Finally, a COVID-19 intelligent auxiliary diagnosis system convenient for doctors is built by using the trained model and the open-source software of Hugging Face.The model achieved 97.98% accuracy on the COVID-QU-Ex Dataset test set, while the parameters and MACs of the model were only 2.23M and 0.33G, which was easy to deploy on hardware devices. Conclusion The intelligent diagnosis system can help doctors to diagnose COVID-19 based on chest radiographs, and improve the accuracy and efficiency of diagnosis.